Selecting a non-negative factorization model for statistical inference on time series of graphs

While non-negative factorization is a popular tool for analyzing non-negative data, model selection procedures for non-negative factorization often lack consideration for stochasticity and its effect on model identification. We consider model selection techniques that can be used to augment existing non-negative factorization algorithms, clarifying the performance of the algorithms for inference on time series of graphs. We demonstrate that our approach reduces the variance of our estimate from non-negative factorization, and is useful for assessing the quality of the estimate. We motivate our approach with singular value decomposition, and illustrate our framework through numerical experiments using real and simulated data.